Don't capture muladds in loop bounds expr

Author: chriselrod

src/vectorizationbase_compat/contract_pass.jl

@@ -207,8 +207,11 @@ function contract!(expr::Expr, ex::Expr, i::Int, mod)
 end
 # contract_pass(x) = x # x will probably be a symbol
 function contract_pass!(expr::Expr, mod = nothing)
-    for (i,ex) ∈ enumerate(expr.args)
-        ex isa Expr || continue
+    i = Core.ifelse(expr.head === :for, 1, 0)
+    while i < length(expr.args)
+        _ex = expr.args[(i+=1)]
+        _ex isa Expr || continue
+        ex::Expr = _ex
         contract!(expr, ex, i, mod)
     end
 end

test/gemm.jl

@@ -71,9 +71,12 @@
         end
     end
     function AmulBavx1!(C, A, B)
-        @avx for m ∈ 1:size(A,1), n ∈ axes(B,2)
+        dM, rM = divrem(size(C,1), 3)
+        dN, rN = divrem(size(C,2), 3)
+        dK, rK = divrem(size(B,1), 3)
+        @avx for m ∈ 1:3*dM + rM, n ∈ 1:3*dN + rN
             Cₘₙ = zero(eltype(C))
-            for k ∈ 1:size(A,2)
+            for k ∈ 1:3*dK + rK
                 Cₘₙ += A[m,k] * B[k,n]
             end
             C[m,n] = Cₘₙ